Skip to contentSkip to conversation

cgroup: allow deletion of cgroups containing only dying processes

PATCH New Suren Baghdasaryan 3 participants 19 messages Jan 16, 2020 – Feb 12, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 cgroup: allow deletion of cgroups containing only dying processes

A cgroup containing only dying tasks will be seen as empty when a userspace process reads its cgroup.procs or cgroup.tasks files. It should be safe to delete such a cgroup as it is considered empty. However if one of the dying tasks did not reach cgroup_exit then an attempt to delete the cgroup will fail with EBUSY because cgroup_is_populated() will not consider it empty until all tasks reach cgroup_exit. Such a condition can be triggered when a task consumes large amounts of memory and spends enough time in exit_mm to create delay between the moment it is flagged as PF_EXITING and the moment it reaches cgroup_exit.
Fix this by detecting cgroups containing only dying tasks during cgroup destruction and proceeding with it while postponing the final step of releasing the last reference until the last task reaches cgroup_exit.

Fixes c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations")
---
 include/linux/cgroup-defs.h |  3 ++
 kernel/cgroup/cgroup.c      | 65 +++++++++++++++++++++++++++++++++----
 2 files changed, 61 insertions(+), 7 deletions(-)
include/linux/cgroup-defs.h +3 -0
index 63097cb243cb..f9bcccbac8dd 100644
@@ -71,6 +71,9 @@ enum {
 
 	/* Cgroup is frozen. */
 	CGRP_FROZEN,
+
+	/* Cgroup is dead. */
+	CGRP_DEAD,
 };
 
 /* cgroup_root->flags */
kernel/cgroup/cgroup.c +58 -7

107 lines not on this page · raw message

Signature · 1 line
2.25.0.rc1.283.g88dfdc4193-goog
#1 lore raw

PATCH 2/2 kselftest/cgroup: add cgroup destruction test

Add new test to verify that a cgroup with dead processes can be destroyed. The test spawns a child process which allocates and touches 100MB of RAM to ensure prolonged exit. Subsequently it kills the child, waits until the cgroup containing the child is empty and destroys the cgroup.

---
 tools/testing/selftests/cgroup/test_core.c | 113 +++++++++++++++++++++
 1 file changed, 113 insertions(+)
tools/testing/selftests/cgroup/test_core.c +113 -0
index c5ca669feb2b..2a5242ec1a49 100644
@@ -2,7 +2,10 @@
 
 #include <linux/limits.h>
 #include <sys/types.h>
+#include <sys/mman.h>
+#include <sys/wait.h>
 #include <unistd.h>
+#include <fcntl.h>
 #include <stdio.h>
 #include <errno.h>
 #include <signal.h>
@@ -12,6 +15,115 @@
 #include "../kselftest.h"
 #include "cgroup_util.h"
 
+static int touch_anon(char *buf, size_t size)
+{
+	int fd;
+	char *pos = buf;
+
+	fd = open("/dev/urandom", O_RDONLY);
+	if (fd < 0)
+		return -1;
+
+	while (size > 0) {
+		ssize_t ret = read(fd, pos, size);
+
+		if (ret < 0) {
+			if (errno != EINTR) {
+				close(fd);
+				return -1;
+			}
+		} else {
+			pos += ret;
+			size -= ret;
+		}
+	}
+	close(fd);
+
+	return 0;
+}
+
+static int alloc_and_touch_anon_noexit(const char *cgroup, void *arg)
+{
+	int ppid = getppid();
+	size_t size = (size_t)arg;
+	void *buf;
+
+	buf = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON,
+		   0, 0);
+	if (buf == MAP_FAILED)
+		return -1;
+
+	if (touch_anon((char *)buf, size)) {
+		munmap(buf, size);
+		return -1;
+	}
+
+	while (getppid() == ppid)
+		sleep(1);
+
+	munmap(buf, size);
+	return 0;
+}
+
+/*
+ * Create a child process that allocates and touches 100MB, then waits to be
+ * killed. Wait until the child is attached to the cgroup, kill all processes
+ * in that cgroup and wait until "cgroup.events" is empty. At this point try to
+ * destroy the empty cgroup. The test helps detect race conditions between
+ * dying processes leaving the cgroup and cgroup destruction path.
+ */
+static int test_cgcore_destroy(const char *root)
+{
+	int ret = KSFT_FAIL;
+	char *cg_test = NULL;
+	int child_pid;
+	char buf[PAGE_SIZE];
+
+	cg_test = cg_name(root, "cg_test");
+
+	if (!cg_test)
+		goto cleanup;
+
+	for (int i = 0; i < 10; i++) {
+		if (cg_create(cg_test))
+			goto cleanup;
+
+		child_pid = cg_run_nowait(cg_test, alloc_and_touch_anon_noexit,
+					  (void *) MB(100));
+
+		if (child_pid < 0)
+			goto cleanup;
+
+		/* wait for the child to enter cgroup */
+		if (cg_wait_for_proc_count(cg_test, 1))
+			goto cleanup;
+
+		if (cg_killall(cg_test))
+			goto cleanup;
+
+		/* wait for cgroup to be empty */
+		while (1) {
+			if (cg_read(cg_test, "cgroup.procs", buf, sizeof(buf)))
+				goto cleanup;
+			if (buf[0] == '\0')
+				break;
+			usleep(1000);
+		}
+
+		if (rmdir(cg_test))
+			goto cleanup;
+
+		if (waitpid(child_pid, NULL, 0) < 0)
+			goto cleanup;
+	}
+	ret = KSFT_PASS;
+cleanup:
+	if (cg_test)
+		cg_destroy(cg_test);
+	free(cg_test);
+	return ret;
+}
+
 /*
  * A(0) - B(0) - C(1)
  *        \ D(0)
@@ -512,6 +624,7 @@ struct corecg_test {
 	T(test_cgcore_populated),
 	T(test_cgcore_proc_migration),
 	T(test_cgcore_thread_migration),
+	T(test_cgcore_destroy),
 };
 #undef T
 
Signature · 1 line
2.25.0.rc1.283.g88dfdc4193-goog
maintainer
#2 lore raw

cover cgroup: Iterate tasks that did not finish do_exit()

In reply to Suren Baghdasaryan

Hi,
I was looking into the issue and came up with an alternative solution that changes task iteration to be consistent with cgroup_is_populated() check and moving the responsibility to check PF_EXITING on the consumers of iterator API.

I haven't check your approach thoroughly, however, it appears to me it complicates (already non-trivial) cgroup destruction path. I ran your selftest on the iterators approach and it proved working.

Michal Koutný (2):
cgroup: Unify css_set task lists
cgroup: Iterate tasks that did not finish do_exit()

Suren Baghdasaryan (1):
kselftest/cgroup: add cgroup destruction test

 include/linux/cgroup-defs.h                |  15 ++-
 include/linux/cgroup.h                     |   4 +-
 kernel/cgroup/cgroup.c                     |  86 ++++++++--------
 kernel/cgroup/debug.c                      |  16 ++-
 tools/testing/selftests/cgroup/test_core.c | 113 +++++++++++++++++++++
 5 files changed, 176 insertions(+), 58 deletions(-)
Signature · 1 line
2.24.1
maintainer
#3 lore raw

PATCH 1/3 cgroup: Unify css_set task lists

We track tasks of css_set in three different lists. The iterators unnecessarily process each list specially.
Use an array of lists and simplify the iterator code. This is refactoring with no intended functional change.

---
 include/linux/cgroup-defs.h | 15 ++++---
 include/linux/cgroup.h      |  4 +-
 kernel/cgroup/cgroup.c      | 81 +++++++++++++++++++------------------
 kernel/cgroup/debug.c       | 16 ++++----
 4 files changed, 60 insertions(+), 56 deletions(-)
include/linux/cgroup-defs.h +10 -5
index 63097cb243cb..e474b5a67438 100644
@@ -180,6 +180,13 @@ struct cgroup_subsys_state {
 	struct cgroup_subsys_state *parent;
 };
 
+enum css_set_task_list {
+	CSS_SET_TASKS = 0,
+	CSS_SET_TASKS_MG,
+	CSS_SET_TASKS_DYING,
+	CSS_SET_TASKS_COUNT,
+};
+
 /*
  * A css_set is a structure holding pointers to a set of
  * cgroup_subsys_state objects. This saves space in the task struct
@@ -214,14 +221,12 @@ struct css_set {
 
 	/*
 	 * Lists running through all tasks using this cgroup group.
-	 * mg_tasks lists tasks which belong to this cset but are in the
+	 * CSS_SET_TASKS_MG lists tasks which belong to this cset but are in the
 	 * process of being migrated out or in.  Protected by
 	 * css_set_rwsem, but, during migration, once tasks are moved to
-	 * mg_tasks, it can be read safely while holding cgroup_mutex.
+	 * CSS_SET_TASKS_MG, it can be read safely while holding cgroup_mutex.
 	 */
-	struct list_head tasks;
-	struct list_head mg_tasks;
-	struct list_head dying_tasks;
+	struct list_head tasks[CSS_SET_TASKS_COUNT];
 
 	/* all css_task_iters currently walking this cset */
 	struct list_head task_iters;
include/linux/cgroup.h +1 -3
index d7ddebd0cdec..284e6b4b43cc 100644
@@ -58,10 +58,8 @@ struct css_task_iter {
 	struct list_head		*tcset_head;
 
 	struct list_head		*task_pos;
-	struct list_head		*tasks_head;
-	struct list_head		*mg_tasks_head;
-	struct list_head		*dying_tasks_head;
 
+	enum css_set_task_list		cur_list;
 	struct css_set			*cur_cset;
 	struct css_set			*cur_dcset;
 	struct task_struct		*cur_task;
kernel/cgroup/cgroup.c +42 -39

188 lines not on this page · raw message

kernel/cgroup/debug.c +7 -9
index 80aa3f027ac3..26b4908600f7 100644
@@ -124,6 +124,7 @@ static int cgroup_css_links_read(struct seq_file *seq, void *v)
 		struct task_struct *task;
 		int count = 0;
 		int refcnt = refcount_read(&cset->refcount);
+		enum css_set_task_list ls;
 
 		/*
 		 * Print out the proc_cset and threaded_cset relationship
@@ -161,17 +162,14 @@ static int cgroup_css_links_read(struct seq_file *seq, void *v)
 		}
 		seq_puts(seq, "\n");
 
-		list_for_each_entry(task, &cset->tasks, cg_list) {
-			if (count++ <= MAX_TASKS_SHOWN_PER_CSS)
-				seq_printf(seq, "  task %d\n",
-					   task_pid_vnr(task));
+		for (ls = CSS_SET_TASKS; ls < CSS_SET_TASKS_DYING; ++ls) {
+			list_for_each_entry(task, &cset->tasks[ls], cg_list) {
+				if (count++ <= MAX_TASKS_SHOWN_PER_CSS)
+					seq_printf(seq, "  task %d\n",
+						   task_pid_vnr(task));
+			}
 		}
 
-		list_for_each_entry(task, &cset->mg_tasks, cg_list) {
-			if (count++ <= MAX_TASKS_SHOWN_PER_CSS)
-				seq_printf(seq, "  task %d\n",
-					   task_pid_vnr(task));
-		}
 		/* show # of overflowed tasks */
 		if (count > MAX_TASKS_SHOWN_PER_CSS)
 			seq_printf(seq, "  ... (%d)\n",
Signature · 1 line
2.24.1
maintainer
#4 lore raw

Re: [PATCH 1/3] cgroup: Unify css_set task lists

Quoted 5 lines · Michal Koutný

We track tasks of css_set in three different lists. The iterators unnecessarily process each list specially.
Use an array of lists and simplify the iterator code. This is refactoring with no intended functional change.

Quoted 7 lines
---
 include/linux/cgroup-defs.h | 15 ++++---
 include/linux/cgroup.h      |  4 +-
 kernel/cgroup/cgroup.c      | 81 +++++++++++++++++++------------------
 kernel/cgroup/debug.c       | 16 ++++----
 4 files changed, 60 insertions(+), 56 deletions(-)

So, I get the urge to move the lists into an array and thought about doing that while adding the third list for sure. However, it does make code paths which don't walk all lists wordier and the code path is already tricky like hell. Given that there aren't that many places which loop over the lists, if you really wanna clean it up, maybe add an interator over the lists so that both parts of code can look lean?

Thanks.

Signature · 1 line
tejun
maintainer
#5 lore raw

PATCH 2/3 cgroup: Iterate tasks that did not finish do_exit()

In reply to Michal Koutný

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Fixes c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations")
---
 kernel/cgroup/cgroup.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
kernel/cgroup/cgroup.c +3 -2
index b56283e13491..132d258e7172 100644
@@ -4492,11 +4492,12 @@ static void css_task_iter_advance(struct css_task_iter *it)
 			goto repeat;
 
 		/* and dying leaders w/o live member threads */
-		if (!atomic_read(&task->signal->live))
+		if (it->cur_list == CSS_SET_TASKS_DYING &&
+		    !atomic_read(&task->signal->live))
 			goto repeat;
 	} else {
 		/* skip all dying ones */
-		if (task->flags & PF_EXITING)
+		if (it->cur_list == CSS_SET_TASKS_DYING)
 			goto repeat;
 	}
 }
Signature · 1 line
2.24.1
maintainer
#6 lore raw

Re: [PATCH 2/3] cgroup: Iterate tasks that did not finish do_exit()

Quoted 5 lines · Michal Koutný

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Quoted 7 lines

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Yeah, this looks fine to me. Any chance you can order this before the clean up so that we can mark it for -stable.

Thanks.

Signature · 1 line
tejun
author Tested-by
#7 lore raw

Re: [PATCH 2/3] cgroup: Iterate tasks that did not finish do_exit()

Quoted 7 lines · Tejun Heo
Quoted 5 lines · Michal Koutný

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Quoted 7 lines
Quoted 7 lines

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Quoted 4 lines

Yeah, this looks fine to me. Any chance you can order this before the clean up so that we can mark it for -stable.

+1 for reordering. Makes it easier to backport.
Thanks,
Suren.

Quoted 8 lines

Thanks.

--
tejun

--
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]

maintainer
#8 lore raw

Re: [PATCH 2/3] cgroup: Iterate tasks that did not finish do_exit()

Quoted 1 line · Suren Baghdasaryan

Thanks.

Quoted 3 lines
Quoted 2 lines

Yeah, this looks fine to me. Any chance you can order this before the clean up so that we can mark it for -stable.

+1 for reordering. Makes it easier to backport.

The grounds still need to be prepared for css_task_iter to store additional information. Let me see how the preceding changes can be minimized.

Michal

maintainer
#9 lore raw

v2 cover cgroup: Iterate tasks that did not finish do_exit()

Hi,
thanks for your feedback. I'm sending updated solution that changes task iteration to be consistent with cgroup_is_populated() check and moving the responsibility to check PF_EXITING on the consumers of iterator API.

Changes from v1
- v1: https://lore.kernel.org/lkml/[email protected]/
- Swap the fixing and refactoring patch (in order to make the fix easier for backport)
- Don't introduce tasks lists array because of unnecessarily long access code
- Fix comment in selftest
- I leave the CC:stable on discretion of the maintainer

Michal Koutný (2):
cgroup: Iterate tasks that did not finish do_exit()
cgroup: Clean up css_set task traversal

Suren Baghdasaryan (1):
kselftest/cgroup: add cgroup destruction test

 include/linux/cgroup.h                     |   4 +-
 kernel/cgroup/cgroup.c                     |  60 ++++++-----
 tools/testing/selftests/cgroup/test_core.c | 113 +++++++++++++++++++++
 3 files changed, 146 insertions(+), 31 deletions(-)
Signature · 7 lines
2.24.1


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
maintainer
#10 lore raw

PATCH v2 1/3 cgroup: Iterate tasks that did not finish do_exit()

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Fixes c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations")
---
 include/linux/cgroup.h |  1 +
 kernel/cgroup/cgroup.c | 23 ++++++++++++++++-------
 2 files changed, 17 insertions(+), 7 deletions(-)
include/linux/cgroup.h +1 -0
index d7ddebd0cdec..e75d2191226b 100644
@@ -62,6 +62,7 @@ struct css_task_iter {
 	struct list_head		*mg_tasks_head;
 	struct list_head		*dying_tasks_head;
 
+	struct list_head		*cur_tasks_head;
 	struct css_set			*cur_cset;
 	struct css_set			*cur_dcset;
 	struct task_struct		*cur_task;
kernel/cgroup/cgroup.c +16 -7
index 735af8f15f95..a6e3619e013b 100644
@@ -4404,12 +4404,16 @@ static void css_task_iter_advance_css_set(struct css_task_iter *it)
 		}
 	} while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
 
-	if (!list_empty(&cset->tasks))
+	if (!list_empty(&cset->tasks)) {
 		it->task_pos = cset->tasks.next;
-	else if (!list_empty(&cset->mg_tasks))
+		it->cur_tasks_head = &cset->tasks;
+	} else if (!list_empty(&cset->mg_tasks)) {
 		it->task_pos = cset->mg_tasks.next;
-	else
+		it->cur_tasks_head = &cset->mg_tasks;
+	} else {
 		it->task_pos = cset->dying_tasks.next;
+		it->cur_tasks_head = &cset->dying_tasks;
+	}
 
 	it->tasks_head = &cset->tasks;
 	it->mg_tasks_head = &cset->mg_tasks;
@@ -4467,10 +4471,14 @@ static void css_task_iter_advance(struct css_task_iter *it)
 		else
 			it->task_pos = it->task_pos->next;
 
-		if (it->task_pos == it->tasks_head)
+		if (it->task_pos == it->tasks_head) {
 			it->task_pos = it->mg_tasks_head->next;
-		if (it->task_pos == it->mg_tasks_head)
+			it->cur_tasks_head = it->mg_tasks_head;
+		}
+		if (it->task_pos == it->mg_tasks_head) {
 			it->task_pos = it->dying_tasks_head->next;
+			it->cur_tasks_head = it->dying_tasks_head;
+		}
 		if (it->task_pos == it->dying_tasks_head)
 			css_task_iter_advance_css_set(it);
 	} else {
@@ -4489,11 +4497,12 @@ static void css_task_iter_advance(struct css_task_iter *it)
 			goto repeat;
 
 		/* and dying leaders w/o live member threads */
-		if (!atomic_read(&task->signal->live))
+		if (it->cur_tasks_head == it->dying_tasks_head &&
+		    !atomic_read(&task->signal->live))
 			goto repeat;
 	} else {
 		/* skip all dying ones */
-		if (task->flags & PF_EXITING)
+		if (it->cur_tasks_head == it->dying_tasks_head)
 			goto repeat;
 	}
 }
Signature · 1 line
2.24.1
author Tested-by on v2 1/3
#11 lore raw

Re: [PATCH v2 1/3] cgroup: Iterate tasks that did not finish do_exit()

Quoted 91 lines · Michal Koutný

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Fixes c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations")
---
 include/linux/cgroup.h |  1 +
 kernel/cgroup/cgroup.c | 23 ++++++++++++++++-------
 2 files changed, 17 insertions(+), 7 deletions(-)
include/linux/cgroup.h +1 -0
index d7ddebd0cdec..e75d2191226b 100644
@@ -62,6 +62,7 @@ struct css_task_iter {
        struct list_head                *mg_tasks_head;
        struct list_head                *dying_tasks_head;
 
+       struct list_head                *cur_tasks_head;
        struct css_set                  *cur_cset;
        struct css_set                  *cur_dcset;
        struct task_struct              *cur_task;
kernel/cgroup/cgroup.c +16 -8
index 735af8f15f95..a6e3619e013b 100644
@@ -4404,12 +4404,16 @@ static void css_task_iter_advance_css_set(struct css_task_iter *it)
                }
        } while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
 
-       if (!list_empty(&cset->tasks))
+       if (!list_empty(&cset->tasks)) {
                it->task_pos = cset->tasks.next;
-       else if (!list_empty(&cset->mg_tasks))
+               it->cur_tasks_head = &cset->tasks;
+       } else if (!list_empty(&cset->mg_tasks)) {
                it->task_pos = cset->mg_tasks.next;
-       else
+               it->cur_tasks_head = &cset->mg_tasks;
+       } else {
                it->task_pos = cset->dying_tasks.next;
+               it->cur_tasks_head = &cset->dying_tasks;
+       }
 
        it->tasks_head = &cset->tasks;
        it->mg_tasks_head = &cset->mg_tasks;
@@ -4467,10 +4471,14 @@ static void css_task_iter_advance(struct css_task_iter *it)
                else
                        it->task_pos = it->task_pos->next;
 
-               if (it->task_pos == it->tasks_head)
+               if (it->task_pos == it->tasks_head) {
                        it->task_pos = it->mg_tasks_head->next;
-               if (it->task_pos == it->mg_tasks_head)
+                       it->cur_tasks_head = it->mg_tasks_head;
+               }
+               if (it->task_pos == it->mg_tasks_head) {
                        it->task_pos = it->dying_tasks_head->next;
+                       it->cur_tasks_head = it->dying_tasks_head;
+               }
                if (it->task_pos == it->dying_tasks_head)
                        css_task_iter_advance_css_set(it);
        } else {
@@ -4489,11 +4497,12 @@ static void css_task_iter_advance(struct css_task_iter *it)
                        goto repeat;
 
                /* and dying leaders w/o live member threads */
-               if (!atomic_read(&task->signal->live))
+               if (it->cur_tasks_head == it->dying_tasks_head &&
+                   !atomic_read(&task->signal->live))
                        goto repeat;
        } else {
                /* skip all dying ones */
-               if (task->flags & PF_EXITING)
+               if (it->cur_tasks_head == it->dying_tasks_head)
                        goto repeat;
        }
 }
--
Quoted 2 lines

2.24.1

Thanks!

#12 lore raw

Re: [PATCH v2 1/3] cgroup: Iterate tasks that did not finish do_exit()

Quoted 98 lines · Suren Baghdasaryan
Quoted 91 lines · Michal Koutný

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Fixes c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations")
---
 include/linux/cgroup.h |  1 +
 kernel/cgroup/cgroup.c | 23 ++++++++++++++++-------
 2 files changed, 17 insertions(+), 7 deletions(-)
include/linux/cgroup.h +1 -0
index d7ddebd0cdec..e75d2191226b 100644
@@ -62,6 +62,7 @@ struct css_task_iter {
        struct list_head                *mg_tasks_head;
        struct list_head                *dying_tasks_head;
 
+       struct list_head                *cur_tasks_head;
        struct css_set                  *cur_cset;
        struct css_set                  *cur_dcset;
        struct task_struct              *cur_task;
kernel/cgroup/cgroup.c +16 -8
index 735af8f15f95..a6e3619e013b 100644
@@ -4404,12 +4404,16 @@ static void css_task_iter_advance_css_set(struct css_task_iter *it)
                }
        } while (!css_set_populated(cset) && list_empty(&cset->dying_tasks));
 
-       if (!list_empty(&cset->tasks))
+       if (!list_empty(&cset->tasks)) {
                it->task_pos = cset->tasks.next;
-       else if (!list_empty(&cset->mg_tasks))
+               it->cur_tasks_head = &cset->tasks;
+       } else if (!list_empty(&cset->mg_tasks)) {
                it->task_pos = cset->mg_tasks.next;
-       else
+               it->cur_tasks_head = &cset->mg_tasks;
+       } else {
                it->task_pos = cset->dying_tasks.next;
+               it->cur_tasks_head = &cset->dying_tasks;
+       }
 
        it->tasks_head = &cset->tasks;
        it->mg_tasks_head = &cset->mg_tasks;
@@ -4467,10 +4471,14 @@ static void css_task_iter_advance(struct css_task_iter *it)
                else
                        it->task_pos = it->task_pos->next;
 
-               if (it->task_pos == it->tasks_head)
+               if (it->task_pos == it->tasks_head) {
                        it->task_pos = it->mg_tasks_head->next;
-               if (it->task_pos == it->mg_tasks_head)
+                       it->cur_tasks_head = it->mg_tasks_head;
+               }
+               if (it->task_pos == it->mg_tasks_head) {
                        it->task_pos = it->dying_tasks_head->next;
+                       it->cur_tasks_head = it->dying_tasks_head;
+               }
                if (it->task_pos == it->dying_tasks_head)
                        css_task_iter_advance_css_set(it);
        } else {
@@ -4489,11 +4497,12 @@ static void css_task_iter_advance(struct css_task_iter *it)
                        goto repeat;
 
                /* and dying leaders w/o live member threads */
-               if (!atomic_read(&task->signal->live))
+               if (it->cur_tasks_head == it->dying_tasks_head &&
+                   !atomic_read(&task->signal->live))
                        goto repeat;
        } else {
                /* skip all dying ones */
-               if (task->flags & PF_EXITING)
+               if (it->cur_tasks_head == it->dying_tasks_head)
                        goto repeat;
        }
 }
--
Quoted 2 lines

2.24.1

Quoted 1 line

Thanks!

Hi Folks,
If this new version looks good could we get an Ack please? I need to start backporting this fix to Android and would like a confirmation before doing so.
Thanks!

maintainer Applied
#13 lore raw

Re: [PATCH v2 1/3] cgroup: Iterate tasks that did not finish do_exit()

In reply to Michal Koutný
Quoted 13 lines · Michal Koutný

PF_EXITING is set earlier than actual removal from css_set when a task is exitting. This can confuse cgroup.procs readers who see no PF_EXITING tasks, however, rmdir is checking against css_set membership so it can transitionally fail with EBUSY.

Fix this by listing tasks that weren't unlinked from css_set active lists.
It may happen that other users of the task iterator (without CSS_TASK_ITER_PROCS) spot a PF_EXITING task before cgroup_exit(). This is equal to the state before commit c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations") but it may be reviewed later.

Quoted 3 lines
Fixes c03cd7738a83 ("cgroup: Include dying leaders with live threads in PROCS iterations")

Applied to cgroup/for-5.6-fixes.

Thanks.

Signature · 1 line
tejun
maintainer
#14 lore raw

PATCH v2 2/3 cgroup: Clean up css_set task traversal

In reply to Michal Koutný

css_task_iter stores pointer to head of each iterable list, this dates back to commit 0f0a2b4fa621 ("cgroup: reorganize css_task_iter") when we did not store cur_cset. Let us utilize list heads directly in cur_cset and streamline css_task_iter_advance_css_set a bit. This is no intentional function change.

---
 include/linux/cgroup.h |  3 ---
 kernel/cgroup/cgroup.c | 61 +++++++++++++++++++-----------------------
 2 files changed, 28 insertions(+), 36 deletions(-)
include/linux/cgroup.h +0 -3
index e75d2191226b..f1219b927817 100644
@@ -58,9 +58,6 @@ struct css_task_iter {
 	struct list_head		*tcset_head;
 
 	struct list_head		*task_pos;
-	struct list_head		*tasks_head;
-	struct list_head		*mg_tasks_head;
-	struct list_head		*dying_tasks_head;
 
 	struct list_head		*cur_tasks_head;
 	struct css_set			*cur_cset;
kernel/cgroup/cgroup.c +28 -33

93 lines not on this page · raw message

Signature · 1 line
2.24.1
maintainer
#15 lore raw

PATCH v2 3/3 kselftest/cgroup: add cgroup destruction test

In reply to Michal Koutný

From: Suren Baghdasaryan <[email protected]>

Add new test to verify that a cgroup with dead processes can be destroyed. The test spawns a child process which allocates and touches 100MB of RAM to ensure prolonged exit. Subsequently it kills the child, waits until the cgroup containing the child is empty and destroys the cgroup.

---
 tools/testing/selftests/cgroup/test_core.c | 113 +++++++++++++++++++++
 1 file changed, 113 insertions(+)
tools/testing/selftests/cgroup/test_core.c +113 -0
index c5ca669feb2b..76c9dd578ba5 100644
@@ -2,7 +2,10 @@
 
 #include <linux/limits.h>
 #include <sys/types.h>
+#include <sys/mman.h>
+#include <sys/wait.h>
 #include <unistd.h>
+#include <fcntl.h>
 #include <stdio.h>
 #include <errno.h>
 #include <signal.h>
@@ -12,6 +15,115 @@
 #include "../kselftest.h"
 #include "cgroup_util.h"
 
+static int touch_anon(char *buf, size_t size)
+{
+	int fd;
+	char *pos = buf;
+
+	fd = open("/dev/urandom", O_RDONLY);
+	if (fd < 0)
+		return -1;
+
+	while (size > 0) {
+		ssize_t ret = read(fd, pos, size);
+
+		if (ret < 0) {
+			if (errno != EINTR) {
+				close(fd);
+				return -1;
+			}
+		} else {
+			pos += ret;
+			size -= ret;
+		}
+	}
+	close(fd);
+
+	return 0;
+}
+
+static int alloc_and_touch_anon_noexit(const char *cgroup, void *arg)
+{
+	int ppid = getppid();
+	size_t size = (size_t)arg;
+	void *buf;
+
+	buf = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON,
+		   0, 0);
+	if (buf == MAP_FAILED)
+		return -1;
+
+	if (touch_anon((char *)buf, size)) {
+		munmap(buf, size);
+		return -1;
+	}
+
+	while (getppid() == ppid)
+		sleep(1);
+
+	munmap(buf, size);
+	return 0;
+}
+
+/*
+ * Create a child process that allocates and touches 100MB, then waits to be
+ * killed. Wait until the child is attached to the cgroup, kill all processes
+ * in that cgroup and wait until "cgroup.procs" is empty. At this point try to
+ * destroy the empty cgroup. The test helps detect race conditions between
+ * dying processes leaving the cgroup and cgroup destruction path.
+ */
+static int test_cgcore_destroy(const char *root)
+{
+	int ret = KSFT_FAIL;
+	char *cg_test = NULL;
+	int child_pid;
+	char buf[PAGE_SIZE];
+
+	cg_test = cg_name(root, "cg_test");
+
+	if (!cg_test)
+		goto cleanup;
+
+	for (int i = 0; i < 10; i++) {
+		if (cg_create(cg_test))
+			goto cleanup;
+
+		child_pid = cg_run_nowait(cg_test, alloc_and_touch_anon_noexit,
+					  (void *) MB(100));
+
+		if (child_pid < 0)
+			goto cleanup;
+
+		/* wait for the child to enter cgroup */
+		if (cg_wait_for_proc_count(cg_test, 1))
+			goto cleanup;
+
+		if (cg_killall(cg_test))
+			goto cleanup;
+
+		/* wait for cgroup to be empty */
+		while (1) {
+			if (cg_read(cg_test, "cgroup.procs", buf, sizeof(buf)))
+				goto cleanup;
+			if (buf[0] == '\0')
+				break;
+			usleep(1000);
+		}
+
+		if (rmdir(cg_test))
+			goto cleanup;
+
+		if (waitpid(child_pid, NULL, 0) < 0)
+			goto cleanup;
+	}
+	ret = KSFT_PASS;
+cleanup:
+	if (cg_test)
+		cg_destroy(cg_test);
+	free(cg_test);
+	return ret;
+}
+
 /*
  * A(0) - B(0) - C(1)
  *        \ D(0)
@@ -512,6 +624,7 @@ struct corecg_test {
 	T(test_cgcore_populated),
 	T(test_cgcore_proc_migration),
 	T(test_cgcore_thread_migration),
+	T(test_cgcore_destroy),
 };
 #undef T
 
Signature · 1 line
2.24.1
maintainer
#16 lore raw

Re: [PATCH v2 3/3] kselftest/cgroup: add cgroup destruction test

Quoted 11 lines · Michal Koutný

From: Suren Baghdasaryan <[email protected]>

Add new test to verify that a cgroup with dead processes can be destroyed. The test spawns a child process which allocates and touches 100MB of RAM to ensure prolonged exit. Subsequently it kills the child, waits until the cgroup containing the child is empty and destroys the cgroup.

Applied 2-3 to cgroup/for-5.7.

Thanks a lot for working on this!

Signature · 1 line
tejun
maintainer
#17 lore raw

PATCH 3/3 kselftest/cgroup: add cgroup destruction test

In reply to Michal Koutný

From: Suren Baghdasaryan <[email protected]>

Add new test to verify that a cgroup with dead processes can be destroyed. The test spawns a child process which allocates and touches 100MB of RAM to ensure prolonged exit. Subsequently it kills the child, waits until the cgroup containing the child is empty and destroys the cgroup.

---
 tools/testing/selftests/cgroup/test_core.c | 113 +++++++++++++++++++++
 1 file changed, 113 insertions(+)
tools/testing/selftests/cgroup/test_core.c +113 -0
index c5ca669feb2b..2a5242ec1a49 100644
@@ -2,7 +2,10 @@
 
 #include <linux/limits.h>
 #include <sys/types.h>
+#include <sys/mman.h>
+#include <sys/wait.h>
 #include <unistd.h>
+#include <fcntl.h>
 #include <stdio.h>
 #include <errno.h>
 #include <signal.h>
@@ -12,6 +15,115 @@
 #include "../kselftest.h"
 #include "cgroup_util.h"
 
+static int touch_anon(char *buf, size_t size)
+{
+	int fd;
+	char *pos = buf;
+
+	fd = open("/dev/urandom", O_RDONLY);
+	if (fd < 0)
+		return -1;
+
+	while (size > 0) {
+		ssize_t ret = read(fd, pos, size);
+
+		if (ret < 0) {
+			if (errno != EINTR) {
+				close(fd);
+				return -1;
+			}
+		} else {
+			pos += ret;
+			size -= ret;
+		}
+	}
+	close(fd);
+
+	return 0;
+}
+
+static int alloc_and_touch_anon_noexit(const char *cgroup, void *arg)
+{
+	int ppid = getppid();
+	size_t size = (size_t)arg;
+	void *buf;
+
+	buf = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON,
+		   0, 0);
+	if (buf == MAP_FAILED)
+		return -1;
+
+	if (touch_anon((char *)buf, size)) {
+		munmap(buf, size);
+		return -1;
+	}
+
+	while (getppid() == ppid)
+		sleep(1);
+
+	munmap(buf, size);
+	return 0;
+}
+
+/*
+ * Create a child process that allocates and touches 100MB, then waits to be
+ * killed. Wait until the child is attached to the cgroup, kill all processes
+ * in that cgroup and wait until "cgroup.events" is empty. At this point try to
+ * destroy the empty cgroup. The test helps detect race conditions between
+ * dying processes leaving the cgroup and cgroup destruction path.
+ */
+static int test_cgcore_destroy(const char *root)
+{
+	int ret = KSFT_FAIL;
+	char *cg_test = NULL;
+	int child_pid;
+	char buf[PAGE_SIZE];
+
+	cg_test = cg_name(root, "cg_test");
+
+	if (!cg_test)
+		goto cleanup;
+
+	for (int i = 0; i < 10; i++) {
+		if (cg_create(cg_test))
+			goto cleanup;
+
+		child_pid = cg_run_nowait(cg_test, alloc_and_touch_anon_noexit,
+					  (void *) MB(100));
+
+		if (child_pid < 0)
+			goto cleanup;
+
+		/* wait for the child to enter cgroup */
+		if (cg_wait_for_proc_count(cg_test, 1))
+			goto cleanup;
+
+		if (cg_killall(cg_test))
+			goto cleanup;
+
+		/* wait for cgroup to be empty */
+		while (1) {
+			if (cg_read(cg_test, "cgroup.procs", buf, sizeof(buf)))
+				goto cleanup;
+			if (buf[0] == '\0')
+				break;
+			usleep(1000);
+		}
+
+		if (rmdir(cg_test))
+			goto cleanup;
+
+		if (waitpid(child_pid, NULL, 0) < 0)
+			goto cleanup;
+	}
+	ret = KSFT_PASS;
+cleanup:
+	if (cg_test)
+		cg_destroy(cg_test);
+	free(cg_test);
+	return ret;
+}
+
 /*
  * A(0) - B(0) - C(1)
  *        \ D(0)
@@ -512,6 +624,7 @@ struct corecg_test {
 	T(test_cgcore_populated),
 	T(test_cgcore_proc_migration),
 	T(test_cgcore_thread_migration),
+	T(test_cgcore_destroy),
 };
 #undef T
 
Signature · 1 line
2.24.1
#18 lore raw

Re: [PATCH 0/3] cgroup: Iterate tasks that did not finish do_exit()

In reply to Michal Koutný

Hi Michal,

Quoted 5 lines · Michal Koutný

Hi,
I was looking into the issue and came up with an alternative solution that changes task iteration to be consistent with cgroup_is_populated() check and moving the responsibility to check PF_EXITING on the consumers of iterator API.

Yeah, that was my first thought which basically reverts a part of c03cd7738a83. When I first brought up this issue in the other email thread, Tejun's comment was "the right thing to do is allowing destruction of cgroups w/ only
dead processes in it". I assumed, maybe incorrectly, that the desire here is not to include dying processes into cgroup.procs but to allow cgroups with dying processes to be deleted.

To be clear, either way is fine with me since both ways solve the issue and this way the code is definitely simpler. I'll rerun the tests with your patches just to confirm the issue is gone. Thanks!

Quoted 25 lines

I haven't check your approach thoroughly, however, it appears to me it complicates (already non-trivial) cgroup destruction path. I ran your selftest on the iterators approach and it proved working.

Michal Koutný (2):
cgroup: Unify css_set task lists
cgroup: Iterate tasks that did not finish do_exit()

Suren Baghdasaryan (1):
kselftest/cgroup: add cgroup destruction test

 include/linux/cgroup-defs.h                |  15 ++-
 include/linux/cgroup.h                     |   4 +-
 kernel/cgroup/cgroup.c                     |  86 ++++++++--------
 kernel/cgroup/debug.c                      |  16 ++-
 tools/testing/selftests/cgroup/test_core.c | 113 +++++++++++++++++++++
 5 files changed, 176 insertions(+), 58 deletions(-)

--
2.24.1

--
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].